Microsoft is dropping a family of seven new AI models, the company announced on Tuesday during the opening keynote of Build, its annual developer conference.
The blockbuster highlight is the 35-billion-parameter MAI-Thinking-1, which Microsoft AI lead Mustafa Suleyman described onstage during the keynote as Microsoft’s “first reasoning model,” and said that independent early testers “prefer it in overall quality, side-by-side, versus [Anthropic’s Claude] Sonnet 4.6.” MAI-Thinking-1 also scored 97% on the AIME benchmark, which measures advanced mathematical and problem-solving abilities, and “most importantly of all,” a 53% on SWE Bench Pro, which measures the ability of AI agents to handle complex coding tasks. Anthropic’s Claude Opus 4.6 currently scores at 51.9%, but OpenAI’s GPT-5.4 has achieved a 59.1% score, according to data from Scale Labs, the model performance tracking division of Scale AI.
The big selling point was that MAI-Image-1 was trained “entirely from the bottom,” as Suleyman put it. That is, it was trained solely to be an outstanding reasoning model across a range of possible tasks, rather than to perform well on specific benchmark tests. It was also trained “with absolutely zero distillation,” according to Suleyman. “Distillation” is AI industry jargon for using another model, perhaps one built by a rival company, to train a new one; in that way, models can become a kind of scaffolding for other developers to use and build on.
Microsoft’s pitch to enterprise customers basically seems to be that distillation can lead to trouble down the road, since there’s so much uncertainty and legal controversy around the matter of sourcing training data and potential copyright infringement. The more clearly you can track the origin of the data within the AI tools you’re using, the better. MAI-Image-1 “is created with an enterprise-grade, clean, and commercially licensed data lineage, that means you can put it into production in a very trustworthy way with complete confidence.” We’ll have to wait for a more thorough explanation of exactly how Microsoft licensed all the training data for MAI-Thinking-1, but for the time being, that’s how it’s trying to sell the model to businesses.
The other models which Suleyman introduced during the Build keynote include MAI-Image-2.5 and MAI-Image-2.5 Flash, image-generating models which “deliver a step-change in quality” (at the time of this writing, MAI-Image-2.5 is in the number three spot on the Arena.AI text-to-image model scoreboard, just behind Google’s Nano Banana 2); MAI-Transcribe-1.5, which Suleyman said is “the best transcription model in the world”; speech-generation models MAI-Voice-2 and Mai-Voice-2 Flash; and MAI-Code-1-Flash for generating code.
Taken together, it’s the biggest AI product announcement from Microsoft since it debuted the first models it had developed entirely in-house, MAI-Voice-1 and MAI-Voice-1-preview, back in August. The company’s AI efforts had leaned heavily on its partnership with OpenAI before then, but the two companies have since gradually drifted apart.
Its early investments in OpenAI, however, helped Microsoft gain some traction in the early days of the AI boom; it’s one of the few old-school tech giants that has successfully managed to become a leader in that race, while others like Apple and IBM have lagged behind. It’s branded its AI ambitions under the slogan “humanist superintelligence,” through which it’s tried to position itself as a technology company that always keeps human interests and agency front-and-center: “We want to both explore and prioritize how the most advanced forms of AI can keep humanity in control while at the same time accelerating our path towards tackling our most pressing global challenges,” Suleyman wrote in a company blog post in November.
Read the full article here
